openstack
/
murano-apps
Code Issues Proposed changes
murano-apps/Docker/Kubernetes/README.rst

11 KiB
Raw Blame History

Google Kubernetes for Murano

Packages in this folder are required to deploy both Google Kubernetes and applications on top of it.

Contents of each folder need to be zipped and uploaded to Murano Catalog.

You will also need to build a proper Ubuntu-based image for Kubernetes. This can be done using diskimage-builder and DIB elements. The image has to be named ubuntu14.04-x64-kubernetes.qcow2

Overview of Kubernetes

Kubernetes is an open-source container manager by Google. It is responsible to schedule, run and manage docker containers into its own clustered setup.

Kubernetes consists of one or more master nodes running Kubernetes API and one or more worker nodes (aka minions) that are used to schedule containers. Containers are aggregated into pods. All containers in single pod are guaranteed to be scheduled to a single node and share common port space. Thus it can be considered as a container co-location.

Pods can be replicated. This is achieved by creation of Replication Controller which creates and maintain fixed number of pod clones. In Murano replica count is a property of KubernetesPod.

For a more in-depth review of Kubernetes please refer to official documentation.

How murano installs Kubernetes

Currently Murano supports setups with only single API node and at least one worker node. API node cannot be used as a worker node.

To establish required network connectivity model for the Kubernetes Murano sets up an overlay network between Kubernetes nodes using Flannel networking. See flannel for more information.

Because IP addresses of containers are in that internal network and not accessible from outside in order to provide public endpoints Murano sets up a third type of nodes: Gateway nodes.

Gateway nodes are connected to both Flannel and OpenStack Neutron networks and serves as a gateway between them. Each gateway node runs HAProxy. When an application deploys all its public endpoints are automatically registered on all gateway nodes. Thus if user chose to have more than one gateway it will usually get several endpoints for the application. Then those endpoints can be registered in physical load balancer or DNS.

KubernetesCluster

This is the main application representing Kubernetes Cluster. It is responsible for deployment of the Kubernetes and its nodes.

The procedure is:

  1. Create VMs for all node types - 1 for Kubernetes API and requested number for worker and gateway nodes.
  2. Join them into etcd cluster. etcd is a distributed key-value storage used by the Kubernetes to store and synchronize cluster state.
  3. Setup Flannel network over etcd cluster. Flannel uses etcd to track network and nodes.
  4. Configure required services on master node.
  5. Configure worker nodes. They will register themselves in master nodes using etcd.
  6. Setup HAProxy on each gateway node. Configure confd to watch etcd to register public ports in HAProxy config file. Each time new Kubernetes service is created it regenerates HAProxy config.

Internally KubernetesCluster contains separate classes for all node types. They all inherit from KubernetesNode that defines the common interface for all nodes. The deployment of each node is split into several methods: deployInstance -> setupEtcd -> setupNode -> removeFromCluster as described above.

KubernetesPod

KubernetesPod represents a single Kubernetes pod with its containers and associated volumes. KubernetesPod provides an implementation of DockerContainerHost interface defined in DockerInterfacesLibrary. Thus each pod can be used as a drop-in replacement for regular Docker host implementation (DockerStandaloneHost).

All pods must have a unique name within single KubernetesCluster (which is selected for each pod).

Thus KubernetesCluster is an aggregation of Docker hosts (pods) which also handles all inter-pod entities (services, endpoints).

KubernetesPod creates Replication Controllers rather than pods. Replication Controller with replica count equal to 1 will result in single pod being created while it is always possible to increase/decrease replica count after deployment. Replica count is specified using replicas input property.

Pods also may have labels to group them (for example into layers etc.)

Kubernetes actions

Both KubernetesCluster and KubernetesPod expose number of actions that can be used by both user (through the dashboard) and automation systems (through API) to perform actions on the deployed applications.

See http://docs.openstack.org/developer/murano/draft/appdev-guide/murano_pl.html#murano-actions and http://docs.openstack.org/developer/murano/specification/index.html#actions-api for more details on actions API.

KubernetesCluster provides the following actions:

  • `scaleNodesUp`: increase the number of worker nodes by 1.
  • `scaleNodesDown`: decrease the number of worker nodes by 1.
  • `scaleGatewaysUp`: increase the number of gateway nodes by 1.
  • `scaleGatewaysDown`: decrease the number of gateway nodes by 1.

KubernetesPod has the following actions:

  • `scalePodUp`: increase the number of pod replicas by 1.
  • `scalePodDown`: decrease the number of pod replicas by 1.

Applications documentation

Documentation for KubernetesCluster application classes

KubernetesCluster

Represents Kubernetes Cluster and is the main class responsible for deploying both Kubernetes and it's nodes.

isAvailable()

Return whether masterNode.isAvailable() or not.

deploy()

Deploy Kubernetes Cluster.

getIp()

Return IP of the masterNode.

createPod(definition, isNew)

Create new Kubernetes Pod. definition is a dict of parameters, defining the pod. isNew is a boolean parameter, telling if the pod should be created or updated.

createReplicationController(definition, isNew)

Create new Replication Controller. definition is a dict of parameters, definition of the pod. isNew is a boolean parameter, telling if the pod should be created or updated.

deleteReplicationController(id)

Calls kubectl delete replicationcontrollers with given id on master node.

deletePods(labels)

Accepts a dict of labels with string-keys and string-values, that would be passed to kubectl delete pod on master node.

createService(applicationName, applicationPorts, podId)
  • applicationName a string holding application's name.
  • applicationPorts list of instances of com.mirantis.docker.ApplicationPort class.
  • podId a string holding a name of the pod.

Check each port in applicationPorts and creates or updates it if the port differs from what it was before (or did not exist). Calls kubectl replace or kubectl create on master node.

deleteServices(applicationName, podId)
  • applicationName a string holding application's name,
  • podId a string holding a name of the pod.

Delete all of the services of a given pod, calling kubectl delete service for each one of them.

scaleRc(rcName, newSize)
  • rnName string holding the name of the RC
  • newSize integer holding the number of replicas.

Call kubectl scale rc on master node, setting number of replicas for a given RC.

scaleNodesUp()

Increase the number of nodes by one ($.nodeCount up to the len($.minionNodes)) and call .deploy(). Can be used as an Action.

scaleGatewaysUp()

Increase the number of gateways by one ($.gatewayCount up to the len($.gatewayNodes)) and call .deploy(). Can be used as an Action.

scaleNodesDown()

Decrease the number of nodes by one ($.nodeCount up to 1) and call .deploy(). Can be used as an Action.

scaleGatewaysUp()

Decrease the number of gateways by one ($.gatewayCount up to 1) and call .deploy(). Can be used as an Action.

KubernetesNode

Base class for all Kubernetes nodes.

getIp(preferFloatingIp)

Return IP address of the instance. If preferFloatingIp is False (default) return first IP address found. Otherwise give preference to floating IP.

deployInstance()

Call .deploy() method of underlying instance.

KubernetesGatewayNode

Kubernetes Gateway Node. Extends KubernetesNode class. All methods in this class are idempotent. This is achieved by memoizing the fact that the function has been called.

deployInstance()

Deploy underlying instance.

setupEtcd()

Add current node to etcd config (by calling etcdctl member add) on master node and start etcd member service on underlying instance.

setupNode()

Set up the node, by first setting up Flannel and then setting up HAProxy load balancer on underlying instance.

removeFromCluster()

Remove current node from etcd cluster and call $.instance.releaseResources(). Also clear up memoized values for deployInstance, setupEtcd, setupNode, allowing you to call these functions again.

KubernetesMasterNode

Kubernetes Master Node. Extends KubernetesNode class. Most methods in this class are idempotent. This is achieved by memoizing the fact that the function has been called.

deployInstance()

Deploy underlying instance.

setupEtcd()

Set up etcd master node config and launch etcd service on master node.

setupNode()

Set up the node. This includes setting up Flannel for master and configuring and launching kube-apiserver, kube-scheduler and kube-controller-manager services on the underlying instance.

isAvailable()

Return whether underlying instance has been deployed.

KubernetesMinionNode

Kubernetes Minion Node. Extends KubernetesNode class. All methods in this class are idempotent. This is achieved by memoizing the fact that the function has been called.

deployInstance()

Deploy underlying instance.

setupEtcd()

Add current node to etcd config (by calling etcdctl member add) on master node and start etcd member service on underlying instance.

setupNode()

Set up the node, by first setting up Flannel and then joining the minion into the cluster. If dockerRegistry or dockerMirror are supplied for underlying cluster, those are appended to the list of docker parameters. If gcloudKey is supplied for underlying cluster, then current node attempts to login to google cloud registry. Afterwards restart docker and configure and launch kubelet and kube-proxy services

removeFromCluster()

Remove current node from etcd cluster and call $.instance.releaseResources(). Also clear up memoized values for deployInstance, setupEtcd, setupNode, allowing you to call these functions again.